Style Loss is the major player for good results and not Adversarial Loss.

researchmm / AOT-GAN-for-Inpainting

[TVCG'2023] AOT-GAN for High-Resolution Image Inpainting (codebase for image inpainting)

https://arxiv.org/abs/2104.01431

Apache License 2.0

425 stars 68 forks source link

Style Loss is the major player for good results and not Adversarial Loss. #7

Open praeclarumjj3 opened 3 years ago

praeclarumjj3 commented 3 years ago

Hi! Thanks for the excellent codebase!

I ran a few experiments to measure the importance of GAN in the current network. It turns out if we don't use style loss, the results are largely blurry. This makes me wonder about the importance of GAN. Could you help me there?

You can confirm the experiments here: https://github.com/praeclarumjj3/AOT-GAN-Experiments#results-using-the-testing-pconv-mask-dataset-without-style-loss

Also, I fixed the bugs present in the adv_loss as mentioned in #2.

stteovo commented 4 months ago

before I send (img, mask) to the backbone, should I normalize the mentioned "mask" to (0, 1) or (-1, 1). I get differrent results with the two diffrrent way of normalization.